Multi-Max-Margin Support Vector Machine for multi-source human action recognition

We propose a new ensemble-based classifier for multi-source human action recognition called Multi-Max-Margin Support Vector Machine (MMM-SVM). This ensemble method incorporates the decision values of multiple sources and makes an informed final prediction by merging multi-source feature's intrinsic decision strength. Experiments performed on the benchmark IXMAS multi-view dataset (Weinland) demonstrate that the performance of our multi-view system can further improve the accuracy over single view by 3–13% and consistently outperform the direct-concatenation method. We further apply this ensemble technique for combining the decision values of contextual and motion information in the UCF Sports dataset (Liu, 2009) and the results are comparable to the state-of-the-art, which exhibits our algorithm's potential for further extension in other areas of feature fusion problems.